The transition from ad hoc prompting to repeatable AI workflows presents a significant challenge for developers. While prompting large language models like Claude or ChatGPT offers speed, the results are often inconsistent and difficult to reproduce. Conversely, building complex workflows in Python ensures reliability but often sacrifices the flexibility needed for creative exploration. To bridge this gap, a new approach utilizes Claude Code skills, which combine natural language instructions with structured scripts to create consistent yet adaptable processes. This methodology was demonstrated through a case study involving virtual customer research, specifically simulating LLM personas for qualitative interviews. Traditional methods of assigning a persona, such as instructing an AI to act as a twenty-five-year-old skincare enthusiast, often fail in production. Responses tend to drift over time, answers anchor to previous interactions, and the resulting data lacks the necessary diversity for accurate research. Simply refining prompts does not solve these structural issues; the workflow itself requires standardization. The solution lies in converting fragile, multi-step prompting sequences into a reusable Claude Code skill. Instead of manually reconfiguring panel setups and persona details, a single command like /persona generate 10 Gen Z skincare shoppers in the US initiates the entire process. Behind this command, the skill handles panel design, persona generation, validation, and output packaging. A critical innovation in this workflow is treating personas as structured data objects rather than simple text blocks. A naive approach relies on vague conversational instructions that allow character attributes to drift during a conversation. By defining personas as JSON objects with pinned attributes, the workflow maintains stability, allowing the same panel to be reloaded for different concept tests. Furthermore, the skill enforces diversity by defining the desired attitudinal mix before generation. This ensures a balanced panel of segments, such as budget-conscious shoppers versus trend chasers, and validates the distribution afterward. Each interview runs in an isolated context to prevent cross-contamination of answers between different personas. The choice of a Claude Code skill over a Python library or a standard prompt offers distinct advantages. Libraries like Microsoft's TinyTroupe provide similar functionality but require separate API keys, adding cost and friction during experimentation. Skills, however, run within the existing subscription, scaling without extra charges. They also allow parameters to be passed as natural language rather than rigid function signatures. Additionally, the SKILL.md file acts as a guardrail, embedding rules for structure and diversity into the workflow skeleton. This approach is not suitable for fully deterministic pipelines or tasks requiring strict regulatory audit trails, where pure code is preferable. However, for workflows needing structure without the rigidity of full hardcoding, skills are ideal. They allow developers to mix the flexible judgment of the LLM with deterministic Python scripts for validation and aggregation. In the demonstrated case, Python code handled panel diversity checks while the LLM managed the creative generation, creating a robust system for virtual customer interviews. The full implementation, including the SKILL.md file, Python scripts, and a detailed demo, is available on GitHub under the project name claude-persona. This case study highlights how Claude Code skills can solve the tension between the instability of ad hoc prompting and the inflexibility of traditional coding, offering a viable path for scalable, repeatable AI applications.